What does U-Net predict in Stable Diffusion at each step?

In the Stable Diffusion model, at each U-Net step, the network predicts the noise that was added to the image at that stage of the diffusion process. Specifically, U-Net is trained to reconstruct the original image from the noisy version, i.e., it predicts the noise component that needs to be subtracted to obtain a cleaner image.

The process can be described as:

1. A noisy image at the current step is input.
2. U-Net predicts the noise present in this image.
3. The predicted noise is used to update the image, reducing the noise level.

Thus, U-Net plays a key role in the reverse diffusion process, gradually restoring a detailed image from random noise.